Overview
On October 22, at the 2026 iFlytek 1024 Global Developer Festival in Beijing, iFlytek released Spark X2.6, a large model trained entirely on domestic compute. Using a 293B-A30B MoE architecture, it supports a one-million-token context and, for the first time on fully domestic compute, breaks through in code and long-horizon agent tasks.
Underpinning training is the Feixing domestic ten-thousand-card compute base built with Huawei on the Ascend ecosystem, capable of training trillion-parameter models. Chairman Liu Qingfeng said only models truly trained on domestic compute are genuinely self-controllable.
Background & Interpretation
1. Technical Background and Evolution
Model competition is shifting from parameter scale to compute autonomy and deployment efficiency. With US export controls tightening, domestic-compute training has become mandatory. iFlytek’s dynamic parallel-training lifted ultra-long-context training efficiency 4x; on the Feixing-II domestic chip, inference efficiency rose 4.5x and KV Cache hit rate from 63% to 95.7%.
2. Core Drivers and Underlying Mechanism
Full-stack autonomy means the “root” is domestic — a closed loop from compute and frameworks to models and applications. The cloud-edge architecture lets Spark X2.6 run pure on-device, cloud-edge collaborative or pure cloud, splitting tasks by difficulty, data-security needs and capability. Spark X2.5-4B open-sourced one month drew 2.56 million downloads; on-device agent coverage rose from 12.8% to over 40%.
3. Market Structure and Sector Impact
At the festival, an all-domestic AI chain appeared: AI unmanned stores, AI TCM clinics, AI glasses. iFlytek’s robot brain platform 3.0 is integrated into 500+ robot firms and 100,000+ devices; 90% of top domestic robot makers are its clients. The domestic “compute + model + terminal” chain is sharpening.
Implications & Outlook
Near term, full-stack domestic solutions penetrate government, finance and healthcare where data security is sensitive; medium term, the domestic-compute efficiency gap narrows and MoE with cloud-edge coordination becomes mainstream. Key variables: domestic-chip memory-bandwidth breakthroughs, ten-thousand-card cluster utilization, and on-device model ecosystems.
Takeaways for Industry Participants
- Government and enterprise users: prioritize full-stack domestic options for data-sensitive scenarios.
- Developers: tap on-device models and cloud-edge routing to cut token costs.
- Supply chain: opportunities in domestic-compute ecosystem software-hardware coordination (frameworks, inference optimization, industry models).
This column compiles industry information and shares technical perspectives; it does not constitute investment advice.